AACRNet: Ambiguity-aware change refinement network for remote sensing images change detection

Abstract

Change detection is a fundamental task in remote sensing image analysis. Despite the promising results achieved by recent deep learning-based methods, existing approaches still struggle with ambiguous cases, particularly in scenarios with high visual similarity to complex background noise. Such visual ambiguity often results in semantic confusion and erroneous predictions. However, few methods have explicitly addressed this challenge. In this paper, we propose an ambiguity-aware change refinement network (AACRNet) to tackle this issue. Specifically, a Siamese backbone network is utilized to extract multi-level feature representations from bitemporal images. At each hierarchical level, a temporal difference fusion module (TDFM) is introduced to dynamically establish spatiotemporal relationships, capture global change information, and generate difference features. To mitigate semantic discrepancies between adjacent hierarchical difference features and efficiently locate key change regions, we incorporate a semantic aggregation module (SAM) to integrate global semantic information. Furthermore, we propose a focusing refinement module (FRM) to explicitly model ambiguous regions, introducing a novel strategy that exploits global semantic information to extract and distinguish interference signals within these ambiguous regions. By progressively refining cross-layer features, the FRM enhances semantic consistency and reduces false predictions in such areas. Extensive experiments on the WHU-CD, LEVIR-CD, and SYSU-CD datasets indicate that AACRNet outperforms other state-of-the-art (SOTA) change detection methods, achieving superior accuracy and robustness.